Codebind

Codebind turns an ordinary IPython session into a durable model harness with one tool: an IPython cell executed in the active session. IPython owns execution, namespace, history, magics, tracebacks, and rich display; Codebind owns conversation state and model orchestration.

Installation

uvx codebind

Or install it with any Python package installer:

pip install codebind

Configuration

Codebind reads its private model choice from $XDG_CONFIG_HOME/codebind/configuration.json, or ~/.config/codebind/configuration.json when XDG_CONFIG_HOME is unset:

{
  "model": "openai/gpt-5.6-luna",
  "parameters": {
    "reasoning_effort": "medium"
  }
}

Provider credentials use the same XDG directory in models.json. Keep both files readable only by their owner. Loading Codebind does not add chat, model, models, Models, or any other variable to the IPython namespace.

Terminal IPython

Start a new conversation:

uvx codebind

Ask through ordinary IPython magic syntax:

%%question
Inspect this project and tell me what to implement first.

Terminal conversations remain live for the current IPython process. Durable automatic resumption belongs to notebooks, where the notebook file provides an unambiguous conversation identity.

JupyterLab

uvx --from jupyterlab --with codebind --with 'nbconvert[webpdf]' jupyter lab

When running an unpublished local checkout, expose both its editable Python source and its current prebuilt frontend:

JUPYTER_PATH=/path/to/codebind/data/share/jupyter \
uvx --refresh --from jupyterlab --with-editable /path/to/codebind \
  --with 'nbconvert[webpdf]' jupyter-lab

Load Codebind in a notebook:

%load_ext codebind

That is the complete setup. Question is a native toolbar toggle. Turning it on converts the selected cell into a Question and automatically makes each newly created user cell a Question until the toggle is turned off. Codebind-generated instruction, tool, and answer cells are never converted. Write ordinary Markdown and press Shift+Enter.

Loading the extension adds a locked Markdown cell containing Codebind's packaged instructions. That exact text becomes the conversation's immutable system message and remains stable when the server or kernel restarts.

The whole notebook is model context. Before each Question, Codebind takes a canonical snapshot of ordinary Markdown, raw, and code cells, including visible text and image outputs but never the live Python namespace. Images displayed by the model's IPython tool are included in that tool's result as well. PNG, JPEG, WebP, GIF, and SVG renditions are prepared as model-visible images; unsupported or oversized images are reported as omissions rather than silently discarded. The first turn records the snapshot; later turns append only new, changed, removed, or reordered cells. Questions, tool calls, tool results, and answers already present in the conversation ledger are not duplicated. Image bytes are carried in the relevant message, while notebook snapshots contain only stable image descriptors. This append-only representation keeps the prior model prefix unchanged for provider caching.

Codebind stores the complete LangChain message, notebook-context, and turn ledger in notebook metadata. Reopening the notebook, restarting its kernel, and loading the extension restores the exact accumulated context automatically. Each assistant Markdown segment is buffered and inserted as one complete native cell as soon as that segment finishes, before a following tool call is complete. Tool executions and assistant answers are always appended to the notebook end without changing the user's current selection or scroll position. Tool outputs are collapsed by default and remain expandable through JupyterLab's native output control.

The conversation ledger validates its structure and every assistant tool call against its single matching result, without a version gate. Interrupted saves finish before a turn is cancelled, so a restart cannot turn a partial write into an orphan tool result. Temporary model transport failures retry the same Question with a fresh connection; repeated WebSocket failures use HTTP. Invalid requests and authentication failures remain visible errors.

An interrupted tool call is closed with an explicit interrupted result before the turn ends. If a provider stream is cancelled or fails, Codebind discards that provider session before the next Question while retaining the notebook conversation and its stable prompt-cache identity.

The instruction cell, sent Question cells, model-authored tool cells, and assistant Markdown cells are non-editable and non-deletable. Draft Question cells remain editable until they are sent. Jupyter's standard interrupt button cancels an active Codebind question.

Question and assistant Markdown supports $...$, $$...$$, \(...\), and \[...\] through JupyterLab's native MathJax renderer.

Other IPython frontends retain standard MIME display, syntax-highlighted code, Markdown, stdout, tracebacks, rich results, and native IPython history.

ChatGPT account access

Models Provider owns OpenAI account authorization and token refresh. Save the resulting provider-values object in Codebind's XDG models.json; Codebind loads it privately when the extension starts. ChatGPT subscription access is separate from the public, pay-as-you-go OpenAI API and may require compatibility updates when the account protocol changes.